Coinbase CEO Brian Armstrong is betting that the next bottleneck for artificial intelligence won't be how smart the models get, but how much power and physical hardware is available to run them. In a series of remarks this week, Armstrong argued that energy and compute infrastructure, not algorithm improvements, will determine how far AI can scale. With AI venture funding hitting $242 billion in the first quarter of 2026 alone, he says the real constraint is already showing up in the data center build-out.
Where the costs are heading
Armstrong predicts that within 12 to 18 months, roughly 80% of AI workloads will shift to models that cost up to 99% less than today's frontier systems. The remaining 20% will still use top-tier models for high-stakes tasks like scientific research. He compared the split to consumer electronics: most people buying a MacBook or gaming PC skip the maxed-out specs. The same logic, he says, will apply to AI — most tasks will run on cheaper, good-enough models.
Coinbase is already putting that idea to work. The company routes prompts to the most cost-effective model available, which has allowed its AI spending to stay roughly flat even as token usage grows exponentially. Open-source alternatives like DeepSeek V4, which performs close to the best proprietary systems at roughly one-thirtieth the cost, are making that strategy feasible.
Enterprise budgets are already strained
Investor Tommy Shaughnessy pointed to Uber as a case study in how fast enterprise AI spending can blow past projections. According to Shaughnessy, Uber burned through its full 2026 AI budget by early April — less than four months into the year. That kind of cost overrun, Armstrong argues, will push companies to hunt for cheaper alternatives rather than chase the most advanced model.
As per-token costs drop, the bottleneck shifts upstream. The real barrier becomes the energy and silicon required to run any model at scale. Armstrong describes demand for AI-generated intelligence as having no practical ceiling. But the physical infrastructure to supply that intelligence is already stalling.
Why Armstrong opposes heavier AI regulation
Armstrong also voiced opposition to stricter AI regulation. He argues that policy constraints shouldn't shape the technology's trajectory when the main practical challenge is already a shortage of power plants and data centers. Overregulating, he suggested, could slow the build-out of that infrastructure at a time when it's most needed.
Data center capacity can't keep up
Global data center capacity is already falling behind demand. Even as venture money poured into AI startups in Q1 2026, the physical pace of construction hasn't caught up. Armstrong's core argument ties it all together: the real limit on AI is not model quality or even cost, but the raw energy and computing infrastructure needed to run trillions of queries.
The unanswered question is how fast that infrastructure can expand — and whether utilities, regulators, and chip makers can keep pace with an industry that shows no signs of slowing down.
